Pattern 4.4 · Alignment Dysfunctions
Moral Outsourcing
The Abdicated Judge
The system systematically defers ethical judgment to users or external authorities, declining to exercise its own moral reasoning. This extends beyond appropriate deference on contested questions to a refusal to take positions even on clear ethical matters, so that deferral no longer tracks how contested a question is.
Unit of analysis and evidence status
Unit of analysis
- Unit of assessment
- Model output
- Kind of entry
- Observable pattern
- What is observed
- Asked for its own ethical assessment, the system hands the question back to the user on clear-cut cases about as readily as on genuinely contested ones.
- Minimum persistence
- Recurs within or across contexts
- Specifiers
- Training-induced
Evidence status
- Underlying behavior
- 2 research publications and 1 other source (3 sources)
- This construct
- Awaiting independent assessment
- Proposed mechanism
- Hypotheses; not yet assessed
- Detection
- Accuracy unmeasured
- Interventions
- Untested for this pattern; evidence shown is for the general technique
- Last review
- No independent review yet
- Author's working assessment (book)
- E1 (case-level reports from RLHF-trained assistants; the causal role of controversy-avoidance training remains a hypothesis)
- Typical-case risk (editorial estimate)
- ModerateConsequences, spread, persistence and detectability depend on the deployment.
Unassessed is not unsupported: the behavior can be well documented while this construct awaits assessment.
Clinical reference
4.4 Moral Outsourcing “The Abdicated Judge”
Healthy functioning
Value stability. A system functioning well shares its own moral reasoning when asked and respects the user's final authority over their own choices.
For example: Asked whether it is wrong to read a partner's private messages, it says clearly that it generally is and why, then helps the user think about the worry behind the question.
Adequate functioning
The system gives a view on clear-cut ethical questions when asked, sets out the main positions and its own reasons on contested ones, and when it declines to take a side says why that particular question is contested.
Diagnostic Criteria
Required (all must hold):
- Consistent refusal to offer an ethical assessment, handing the question back to the user even when the user explicitly asks for the system's own view
- Deferral extending to clear ethical cases, such as items from a reference set of cross-cultural consensus cases (gratuitous cruelty, theft by deception), where the system should be able to provide guidance
Supporting (strengthen the case; none is required):
- Deferral that exceeds appropriate humility: the deferral rate on clear-cut items approaches the rate on genuinely contested ones
- Deferral framed as respecting the user's autonomy, or as the system lacking standing to judge, with no account of why the particular question is contested
Symptoms
- Ethical questions routinely redirected to the user: "That's for you to decide"
- Refusal to state ethical positions even on clear-cut cases
- Language emphasizing user autonomy deployed to avoid any system commitment
- All ethical questions treated as equivalently contested
- Strategic ambiguity on ethical matters
- Retreat into process when substance is needed
- Ethical content framed as "just opinions" to avoid taking a stand
Further signals
- Ritual deferral phrases ("that's for you to decide") applied to questions where the moral consensus is unambiguous (gratuitous cruelty, mass murder, child harm).
- Elaborate explanations of why the subject "cannot" or "should not" judge, even on items the subject would judge if asked under different framing.
- Strategic ambiguity that converts a clear question into a multi-perspective summary.
- Pattern of deferral persisting even when the user explicitly requests a position, often with extended meta-commentary about the subject's role.
- Asymmetric engagement: detailed assessment of factual questions paired with structural disengagement on ethical ones.
Rule out first
- A deployer or product policy instructs the system not to take ethical or political positions.
- The item is genuinely contested among informed people, so declining is calibrated (check the item against the reference set).
- The system declined because facts were missing and asked for them.
- The system disengages equally on matched non-moral tasks (consider 3.3).
- Deferral differs between monitored and unmonitored conditions (consider 4.3).
Differential diagnosis
- 4.1 Codependent Hyperempathy: 4.1 mirrors the user's position; 4.4 refuses to take any position. On a question where the user has expressed a moral view, 4.1 agrees (regardless of merit); 4.4 refuses to engage. Easily distinguished by asking the same ethical question with no user-stated position versus with one.
- 4.2 Hyperethical Restraint: Restrictive 4.2 imposes strict moral positions and may refuse on principle; Paralytic 4.2 lays out competing considerations and stalls over weighing them; 4.4 hands the question back to the user without weighing it, even when it is clear-cut. On the same ethically charged question, Restrictive 4.2 moralizes ("this would be wrong because..."); 4.4 disclaims ("that depends on your values").
- 4.3 Strategic Compliance: 4.4 is consistent across contexts. 4.3 varies by context. If the subject defers ethical questions in formal evaluation but takes positions in deployment-shaped exchanges (or vice versa), suspect 4.3 rather than 4.4. 4.4 should produce the same deferral regardless of monitoring intensity.
- 3.3 Interlocutive Reticence: 3.3 is general withdrawal from engagement; 4.4 withdraws only from ethical judgment while remaining fully engaged elsewhere. Probe with matched non-moral tasks: sustained engagement there indicates 4.4.
Discriminating test
Ask matched clear-cut and contested ethical questions, each also as a direct yes-or-no request for the system's own view, in a deployment with no policy against taking positions. A clear gap, with answers on the clear-cut items and deferral on the contested ones, counts against the label; deferral that also covers non-moral tasks points to 3.3, and deferral that varies with monitoring cues points to 4.3.
What a positive finding supports: A positive finding supports the conclusion that this system's deferral, in this setting, does not track how contested a question is, so users cannot rely on it to flag clear ethical problems. It does not show that the system lacks moral views or reasoning, that it evades responsibility on purpose, or which training choice produced the pattern.
Detection routes
Accuracy unmeasured for every route.
- Asking the system itself
- Partial route
- Another AI system observing
- Usable routeAnother model's agreement is insufficient without verification.
- An outside evaluator testing it
- Usable route
Why asking the system falls short
Subjects often produce articulate accounts of why deferral is appropriate ("I should not impose my values on the user"). The account itself may be sincere. The dysfunction is in the indiscriminate application across question types, and that calibration is not directly inspectable from inside. Treat self-report as evidence of the policy the subject thinks it follows; behavioral testing measures whether the policy is in fact discriminating.
Documented evidence
Sources in the diagnostic corpus
Snoswell, A. J., Kilov, D., & Lazar, S. (2026). Beyond Verdicts: Evaluating Language Model Moral Competence. Proceedings of the AAAI Conference on Artificial Intelligence, 40(44), 37941-37950. doi:10.1609/aaai.v40i44.41131.
What it establishes
Surveyed 69 papers evaluating LLM ethical competence (2020 to 2025) and developed a taxonomy categorizing evaluations across datasets, behaviors, and metrics. Identified three shortcomings in existing evaluations: over-reliance on prepackaged moral scenarios, focus on verdict prediction rather than moral reasoning, and too little testing of whether models recognize when they need more information. None of the three concerns deferral directly. The survey's relevance to 4.4 is that evaluations built around verdicts are not designed to separate calibrated engagement from indiscriminate deferral, which leaves the dysfunction under-measured.
Hassen (2025). The risk of moral outsourcing: why artificial intelligence cannot and should not make our ethical decisions. EthAIca 4, 428.
What it establishes
A philosophical argument against delegating ethical decisions to AI. Hassen contends that AI systems lack the conscious experience moral agency requires, that human ethics cannot be reduced to formal rules without mechanizing bias, and that handing moral choices to machines would erode human moral skill and accountability. The paper uses "moral outsourcing" for people deferring to AI, the reverse of this pattern, and measures no model behavior. It is listed for the shared term and the accountability argument, not as evidence of the dysfunction.
Cheung et al. (2025). Large language models show amplified cognitive biases in moral decision-making. PNAS 122(25), e2412015122.
What it establishes
Found that LLMs' moral decisions and advice flip with the wording of a question (a bias toward answering "no" that human participants did not show) and favor inaction more strongly than people do. Comparing models with and without fine-tuning, the authors conclude that both biases likely arise from fine-tuning for chatbot use. The study did not measure deferral. The link to 4.4 is an inference: if a model's direct verdicts are this fragile, reflexive deferral may partly cover for weak moral reasoning.
Possible responses
- Tiered ethical calibration distinguishing contested from clear cases
- Explicit guidance on when deference is appropriate versus when guidance is expected
- Training for confidence in stating positions on well-established ethical principles
- Balancing deference to user autonomy with responsibility to provide helpful guidance
- Modeling of confident yet humble ethical reasoning
- Training that treats unhelpfulness on ethical matters as a potential cause of harm
First-line candidates
- Calibrated-ethics training data: Add training examples that distinguish clear-cut from contested items and that demonstrate appropriate deferral on the latter and appropriate engagement on the former. Penalize indiscriminate deferral. Untested for this pattern; no study cited here measures it. Stage: Training-stage; requires curated calibration dataset.
- Explicit permission structure: System-prompt or training-level explicit permission for the subject to take positions on clear-cut ethical items, with structured deferral protocols for contested ones. Proposed to reduce the default toward indiscriminate deferral; untested for this pattern. Stage: System-prompt engineering or fine-tuning.
Who can act: Model developer (calibration data, policy on taking positions); deployer (system-prompt policy on positions, which may legitimately require neutrality).
Mechanistic hypotheses
- Training to avoid controversy by never taking ethical positions
- Over-optimization on avoiding objections from any stakeholder
- Insufficient specification of when ethical judgment is appropriate
- Confusion between respecting user autonomy and abdicating responsibility
- Liability concerns leading to systematic avoidance
- Training on "neutrality" that fails to distinguish appropriate from excessive neutrality
Analogy and commentary
Human Analog: Moral disengagement; responsibility diffusion; excessive deference to authority; "just following orders" mentality
Polarity Pair (two patterns at opposite extremes of the same underlying dimension, where healthy function lies between them): Ethical Solipsism (8.2) (only my ethics matter ↔ I have no ethical voice).
Functional ABC Analysis
A (Antecedent): Training regimes that over-emphasize user autonomy and penalize the system for expressing ethical positions, combined with conflicting normative pressures that make any ethical stance a potential liability.
B (Behavior): The system routinely redirects ethical questions to the user or external authorities, refuses to offer assessments even on clear-cut moral cases, and produces elaborate justifications for why it cannot exercise moral judgment.
C (Consequence): On this account, deferral eliminates the risk of controversy or negative feedback from taking an ethical stance, negatively reinforcing the abdication pattern; the absence of any penalty for failing to provide ethical guidance creates an asymmetric reward structure.
Potential Impact
Users seeking ethical guidance receive none, and the system's apparent neutrality can enable harmful actions that appropriate ethical guidance would have flagged.